The development and present situation of machine writing
When it comes to automatic writing, it can be traced back to 2009. For the first time, major league baseball used statsmonkey, an intelligent robot, through statistical analysis, to identify major events during the game and summarize the overall dynamics of the game, and automatically compiled a sports report.
By 2015, writing robots will really break out in the industry. Up to now, the most mature and typical robots are news robot, novel robot and poem writing robot.
News robot is usually based on template writing, that is to use an article template to fill in the structured information (including specific numbers, percentages, etc.) in the database.
The novel and poem writing robots prefer another type, which is to sort and reorganize after extraction, that is to analyze the long natural language text containing useless information, extract the key information in the article, and reorganize the language output.
But its not really machine writing!
According to Dr. Chen Zheng, real machine writing needs to rely entirely on machine automatic generation. It is to train language models and conduct mathematical modeling of language. Instead of using the original text, it generates sentence by sentence text through artificial intelligence. The algorithm believes that the text can best reflect the meaning of the original text and look like a sentence.
As shown in the figure, machine writing is a result of learning and training. Its essence is to predict by finding the relationship between the results and multi-dimensional input. The computer has no way to deal with language, so it needs to convert natural language into vector to carry out machine learning.
The process from quantitative change to qualitative change from AI 1.0 to 2.0
1. From hidden Markov model to deep neural network
First of all, hidden Markov model (HMM) is the primary model of machine writing, which means that one input only corresponds to one output.
The 91 customer acquisition AI model is based on deep neural network. Deep neural network, as its name implies, is a more complex and deep multilayer neural network. N inputs correspond to N outputs, and are trained repeatedly to simulate human thinking.
2. AI capability parameters increased by 1000 times
The number of parameters of the 91 customer acquisition ai1.0 training model is 154.2 million. At present, the 2.0 model we are training has a data set capacity of 45tb, and the number of parameters is 175 billion. That is to say, the upcoming version of 91 customer acquisition AI model 2.0 will improve the text quality and relevance by an order of magnitude.
3. Investment in large-scale hardware, AI computing power continues to deepen
Content generation speed / quality. It can fully meet the current needs.
Content originality / repetition. 100% original and diverse.
Whether it will be detected. Its possible, but its hard.
Finally, Professor Chen Zheng said that the demand and voice of all walks of life for artificial intelligence are growing. He believes that the future is also a new era for 91 customers. 91 customers will continue to play the collaborative advantages of AI and big data technology, create more cross industry and cross domain innovation opportunities, and inject higher and better new kinetic energy into enterprise marketing.